Nemotron 3.5 Super VL
Nemotron 3.5 Super VL is the newest model from NVIDIA, a multimodal hybrid Mamba2-Transformer Mixture-of-Experts model. From its 124.5B parameters, 120.7B sit in the language model, 2.9B in the multi-token prediction (MTP) layer, and 0.8B in the vision encoder and projector. About 12.8B parameters are active per token: 22 of 512 routed experts in each MoE layer, plus one shared expert.
The network has 88 layers (40 Mamba2, 40 MoE and 8 attention), a hidden size of 4,096, and experts that work in a 1,024-dimensional latent space. Vision comes from a RADIO ViT-H/16 encoder with 32 layers.
Most of the parameter budget is in the routed experts, which makes them the natural place to compress.
CompactifAI in action
CompactifAI, our proprietary framework to compress and optimize models, analyzes the model, compress it and restore to match or even improve the original performance.
The pipeline runs in four main stages: tensor-network profiling, compression, healing and reinforcement learning. The stages are not applied in isolation. Profiling generates several candidate compression configurations, each one gets a short healing run, and the candidate that recovers best moves on to full healing.
Tensor-network profiling
The first stage analyzes the model to find where its parameters carry redundant or low-impact information. The profiler uses tensor-network methods, a family of decompositions with a long history in physics, which we apply at the scale of frontier language models. The profiler is only an analysis tool: it does not change the model's architecture, so the delivered model is a standard transformer.
Compression
Once identified all the potential candidates from the previous stage, the compression removes these parts from the original model. In the case of Nemotron we applied compression in two axes:
Expert pruning. Each of the 40 MoE layers, and the MTP layer, goes from 512 to 256 routed experts.
Intermediate-size reduction. The routed experts' intermediate size drops from 2,688 to 2,432, with a compensation step for the removed dimensions.
The Mamba2 and attention layers, the shared experts, the embeddings, the vision encoder and the projector keep their original sizes. The layer structure, tokenizer and chat template are unchanged, and the MTP layer stays in the model.
Routing is untouched, so each token still goes through 22 experts. That is why the active parameter count falls only from 12.8B to 12.3B, while the total halves. Active counts include the embeddings and output head (11.2B without them for Comet, 11.7B for the base). The BF16 checkpoint is 122.3 GB.
Healing
Compressing alone leaves a damaged model, so we follow it with knowledge distillation from the original Nemotron 3.5 Super VL. The original writes the training responses, and Comet trains toward its token-level output distribution.
The healing data covers three areas:
- Single-turn text: math, science, code, reasoning, instruction following and multilingual chat.
- Multi-turn agentic trajectories: terminal, software-engineering and tool-use sessions.
- Images and video.
The MTP layer is pruned like the main model and distilled together with the main head. Every Comet 60B and Nemotron 3.5 Super VL result below was produced with MTP speculative decoding on, using 5 draft tokens, except NVIDIA's NVFP4 checkpoint (see NVFP4).
Reinforcement Learning
Post-training targets what healing leaves behind. The model undergoes GRPO-style training on several environments to further improve on key areas like multi-turn conversation, multi-step tool use and mitigation of hallucinations, while controlling for chain of thought length and verbosity.
Results
We ran Nemotron 3.5 Super VL and Comet 60B through the same harnesses, prompts, judges and sampling settings, so their scores are directly comparable. Decoding used temperature 1.0 and top_p 0.95 with thinking on and no output-token cap beyond the 256K context (a few harnesses set their own), with BF16 weights and an FP8 KV cache, on 4× H200 for Nemotron 3.5 Super VL and 2× H200 for Comet 60B. For context, the tables also include four other models in BF16, run on the same harnesses with each vendor's recommended sampling settings: Qwen3.8 27B, NVIDIA Nemotron 3 Super (text-only, so it appears only in the text table), Inkling-Small and Mistral Medium 3.5.
Comet beats the original uncompressed model on several of the benchmarks the reinforcement learning stage targeted. It is ahead on all three τ²-Bench domains, most clearly telecom (79.28 against 76.75) and retail (85.09 against 82.79), and also ahead on AutomationBench-AA (33.18 against 31.16). It also hallucinates less: on AA-Omniscience it abstains more instead of guessing, and its hallucination rate is 13.4%, less than half the original's 29.1%. The trade-off is fewer correct answers, so its Omniscience Index is slightly lower (2.83 against 5.08).
Vision and video
Comet stays within 3 points of the original on 7 of the 10 image, document and video scores, and within 5 points on 9. It is ahead on BabyVision. The largest gap is V* (−5.8).
Comet 60B ahead of Nemotron 3.5 Super VL. All models in BF16. Other models ran on the same harnesses with each vendor's recommended sampling settings. — = not run, or the model takes no video input (Inkling-Small and Mistral Medium 3.5).
Reasoning and coding
Comet stays within 3 points of the original on 15 of the 18 text and agentic scores, and within 5 points on 17. It is ahead on all three τ²-Bench domains, AutomationBench-AA, SciCode, SWE-bench Pro and GDP.pdf (mean). The largest gap is GDPval-AA (−5.5).
Comet 60B ahead of Nemotron 3.5 Super VL. All models in BF16. Other models ran on the same harnesses with each vendor's recommended sampling settings. τ²-Bench (average) is the mean of the three domains. For Nemotron 3 Super, Inkling-Small and Mistral Medium 3.5, τ²-Bench is a single run (8 runs averaged for Nemotron 3.5 Super VL and Comet 60B), and Mistral Medium 3.5's SWE-bench Pro is a single run with 34 of 731 issues missing, counted as failures. AA-Omniscience and AutomationBench-AA use the public sets, and GDPval-AA is NeMo Gym's 0–100 rubric score, not Artificial Analysis's Elo: none of these are Artificial Analysis leaderboard numbers. — = not run, or not applicable (GDP.pdf needs image input, and Nemotron 3 Super is text-only).
NVFP4
Comet 60B also comes as a 40.5 GB NVFP4 checkpoint for NVIDIA Blackwell GPUs. It was quantized with NVIDIA Model Optimizer in the same mixed-precision layout as NVIDIA's Nemotron 3.5 Super VL NVFP4 (75.1 GB), then refined with quantization-aware distillation from the original model. We ran both NVFP4 checkpoints on 1× H200 with the same harness; H200 has no native NVFP4 support, so vLLM runs them there with weight-only FP4 (Marlin) kernels. Comet stays within 3 points on 21 of the 28 scores and within 5 points on 26. It is ahead on all three τ²-Bench domains, τ³-Banking and AutomationBench-AA, and level on BabyVision. The largest gap is the AA-Omniscience Index (−13.7).
Comet 60B NVFP4 level with or ahead of Nemotron 3.5 Super VL NVFP4. NVIDIA's NVFP4 checkpoint ran without MTP speculative decoding, as its MTP layer does not load in the vLLM build we used (0.28.1rc1.dev628+g2a02f6efe); MTP affects speed, not scores.
How we evaluated
Most text, coding and agentic benchmarks ran through NVIDIA NeMo Gym, including GPQA Diamond, HLE, AA-Omniscience, AA-LCR, GDP.pdf, SciCode, Terminal-Bench 2.1, SWE-bench Verified, SWE-bench Multilingual, SWE-bench Pro, τ²-Bench, τ³-Banking, GDPval-AA and AutomationBench-AA. MMMU-Pro and BLINK ran on our in-house harness. The remaining image, OCR and document benchmarks ran through VLMEvalKit, following NVIDIA's evaluation protocol with original media files, and Video-MME ran through lmms-eval (2 frames per second, up to 256 frames). Some judges and user simulators differ from NVIDIA's (AA-LCR v1.1 judged by gpt-5.6-luna, and gpt-6-luna as the τ²-Bench and τ³-Banking user), and NVIDIA reports Video-MME with VLMEvalKit, so these scores can differ from NVIDIA's published ones. Single-run benchmarks move by 1–2 points between runs, so we treat differences under about 3 points as noise.
Availability
Comet 60B is available today on Hugging Face as BF16 , FP8 and NVFP4. To explore Multiverse Computing's open-source model releases, visit huggingface.com/MultiverseComputingCAI